The outputs for arbitary images look bad. What datasets can be used to train the network for new images?

[jwsoh]

The outputs for arbitary images look bad. What datasets can be used to train the network for new images?

[JWSoh]

For an arbitrary image, the kernel information is required to get a plausible output. Without the kernel, the default setting is "bicubic" downsampling method.

I don't think using other dataset instead of DIV2K would give considerable performance gain, because DIV2K is a quite good high-resolution natural image dataset.
Of course, if the case is very specific, using other dataset could be better.